-
Notifications
You must be signed in to change notification settings - Fork 0
/
training_cityscapes.py
381 lines (316 loc) · 12.1 KB
/
training_cityscapes.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
# USAGE: python3 training_cityscapes.py --arch deeplabv3+ --traintype parallel --epochs 50 --outpath savedModels
import torch
from torch.utils.data import DataLoader
import torch.nn as nn
import torch.optim as optim
from tqdm import tqdm
import matplotlib.pyplot as plt
import time
import segmentation_models_pytorch as smp
from segtransformer import segformer_mit_b3
from liteseg_model.liteseg import LiteSeg
from dataset_cityscapes import cityscapesLoader
import config
import os
import argparse
import yaml
from addict import Dict
from utils import cross_entropy2d, get_metrics, runningScore
ap = argparse.ArgumentParser()
ap.add_argument('-a', '--arch', default='unet', choices=['unet', 'manet', 'linknet', 'pspnet', 'pan', 'deeplabv3', 'deeplabv3+', 'manet','fpn','segformer-b3', 'liteseg'], help='Choose different semantic segmention architecture')
ap.add_argument('-t', '--traintype', default='single', choices=['single', 'parallel'], help='Choose Parallel if 2 GPUs are available')
ap.add_argument("-e", "--epochs", type=int, help="No of Epochs for training")
ap.add_argument("-o", "--outpath", required=True, help="Model Save path ")
args = vars(ap.parse_args())
os.environ['CUDA_VISIBLE_DEVICES'] = "0,1"
path_data = config.CITYSCAPES_DATASET
train_epochs = args['epochs']
DISTRIBUTED_TRAINING = args['traintype']
print("Number of cuda {}".format(torch.cuda.device_count()))
if torch.cuda.device_count() > 1:
print("Let's use : {}".format(torch.cuda.device_count(), "GPUs!"))
print("Let's Use PARALLEL MODE TRAINING")
DISTRIBUTED_TRAINING = True
else:
print("Let's Use SINGLE MODE TRAINING")
DISTRIBUTED_TRAINING = False
## If there is a "RuntimeError: CUDA out of memory",then change the BATCH_SIZE to some lower number.
if DISTRIBUTED_TRAINING:
BATCH_SIZE = config.BATCH_SIZE
else:
BATCH_SIZE = 4 # Change this number based on your System capacity.Normally,11GB Graphics card this is fine.
OUTPUT_PATH = args['outpath']
train_data = cityscapesLoader(
root = path_data,
split='train'
)
val_data = cityscapesLoader(
root = path_data,
split='val'
)
train_loader = DataLoader(
train_data,
batch_size = BATCH_SIZE,
shuffle=True,
num_workers = config.NUM_WORKERS,
#pin_memory = pin_memory # gave no significant advantage
)
val_loader = DataLoader(
val_data,
batch_size = BATCH_SIZE,
num_workers = config.NUM_WORKERS,
#pin_memory = pin_memory # gave no significant advantage
)
ENCODER = 'resnet50'
ENCODER_WEIGHTS = 'imagenet'
CLASSES = 19
ACTIVATION = 'sigmoid' # could be None for logits or 'softmax2d' for multiclass segmentation
model = None;
if args["arch"] == "pan":
# create segmentation model with pretrained encoder
model = smp.PAN(
encoder_name=ENCODER,
encoder_weights=ENCODER_WEIGHTS,
classes=CLASSES,
activation=ACTIVATION,
)
elif args["arch"] == "unet":
model = smp.Unet(
encoder_name=ENCODER,
encoder_weights=ENCODER_WEIGHTS,
classes=CLASSES,
activation=ACTIVATION,
)
elif args["arch"] == "manet":
model = smp.MAnet(
encoder_name=ENCODER,
encoder_weights=ENCODER_WEIGHTS,
classes=CLASSES,
activation=ACTIVATION,
)
elif args["arch"] == "linknet":
model = smp.Linknet(
encoder_name=ENCODER,
encoder_weights=ENCODER_WEIGHTS,
classes=CLASSES,
activation=ACTIVATION,
)
elif args["arch"] == "pspnet":
model = smp.PSPNet(
encoder_name=ENCODER,
encoder_weights=ENCODER_WEIGHTS,
classes=CLASSES,
activation=ACTIVATION,
)
elif args["arch"] == "deeplabv3":
model = smp.DeepLabV3(
encoder_name=ENCODER,
encoder_depth=5,
encoder_weights=ENCODER_WEIGHTS,
classes=CLASSES,
activation=ACTIVATION,
)
elif args["arch"] == "manet":
model = smp.MANet(
encoder_name=ENCODER,
encoder_weights=ENCODER_WEIGHTS,
classes=CLASSES,
activation=ACTIVATION,
)
elif args["arch"] == "deeplabv3+":
model = smp.DeepLabV3Plus(
encoder_name=ENCODER,
encoder_depth=5,
encoder_weights=ENCODER_WEIGHTS,
classes=CLASSES,
activation=ACTIVATION,
)
elif args["arch"] == "fpn":
model = smp.FPN(
encoder_name=ENCODER,
encoder_weights=ENCODER_WEIGHTS,
classes=CLASSES,
activation=ACTIVATION,
)
elif args["arch"] == "segformer-b3":
model = segformer_mit_b3(in_channels=3, num_classes=CLASSES)
elif args["arch"] == "liteseg":
backbone_network = "mobilenet"
CONFIG = Dict(yaml.load(open("liteseg_model/config/training.yaml"), Loader=yaml.Loader))
model = LiteSeg.build(backbone_network, None, CONFIG, is_train=True, classes=CLASSES)
if DISTRIBUTED_TRAINING:
model = nn.DataParallel(model)
model = model.cuda()
else:
model = model.to(config.DEVICE)
optimizer = optim.Adam(model.parameters(), lr=config.LEARNING_RATE)
def train(train_loader, model, optimizer, epoch_i, epoch_total):
count = 0
# List to cumulate loss during iterations
loss_list = []
for (images, labels) in train_loader:
count += 1
# we used model.eval() below. This is to bring model back to training mood.
model.train()
#pred = None
if DISTRIBUTED_TRAINING:
images = images.cuda()
labels = labels.cuda()
# Model Prediction
#pred = model(images).cuda()
else:
images = images.to(config.DEVICE)
labels = labels.to(config.DEVICE)
#pred = model(images).to(config.DEVICE)
pred = model(images)
# Loss Calculation
loss = cross_entropy2d(pred, labels)
loss_list.append(loss)
# optimiser
optimizer.zero_grad()
loss.backward()
optimizer.step()
# interval to print train statistics
if count % 50 == 0:
fmt_str = "Image: {:d} in epoch: [{:d}/{:d}] and Loss: {:.4f}"
print_str = fmt_str.format(
count,
epoch_i + 1,
epoch_total,
loss.item()
)
print(print_str)
# # break for testing purpose
# if count == 10:
# break
return(loss_list)
def save_model(network, epoch_label):
print("save model: epoch_label = {}".format(epoch_label))
save_filename = '{}_{}.pth'.format(args['arch'], epoch_label)
save_path = os.path.join('./savedModels', save_filename)
if DISTRIBUTED_TRAINING:
torch.save(network.module.state_dict(), save_path)
else:
torch.save(network.state_dict(), save_path)
def validate(val_loader, model, epoch_i):
# tldr: to make layers behave differently during inference (vs training)
model.eval()
# enable calculation of confusion matrix for n_classes = 19
running_metrics_val = runningScore(19)
# empty list to add Accuracy and Jaccard Score Calculations
acc_sh = []
js_sh = []
with torch.no_grad():
for image_num, (val_images, val_labels) in tqdm(enumerate(val_loader)):
if DISTRIBUTED_TRAINING:
val_images = val_images.cuda()
val_labels = val_labels.cuda()
else:
val_images = val_images.to(config.DEVICE)
val_labels = val_labels.to(config.DEVICE)
# Model prediction
val_pred = model(val_images)
# Coverting val_pred from (1, 19, 512, 1024) to (1, 512, 1024)
# considering predictions with highest scores for each pixel among 19 classes
pred = val_pred.data.max(1)[1].cpu().numpy()
gt = val_labels.data.cpu().numpy()
# Updating Mertics
running_metrics_val.update(gt, pred)
sh_metrics = get_metrics(gt.flatten(), pred.flatten())
acc_sh.append(sh_metrics[0])
js_sh.append(sh_metrics[1])
accuracy = sh_metrics[0]
print("sh_metrics[0]: {}".format(sh_metrics[0]))
print("epoch_i: {}".format(epoch_i))
print("config.best_accuracy: {}".format(config.best_accuracy))
if (epoch_i %10 == 0 and accuracy > config.best_accuracy):
config.best_accuracy = accuracy
save_model(model, epoch_i)
# # break for testing purpose
# if image_num == 10:
# break
score = running_metrics_val.get_scores()
running_metrics_val.reset()
acc_s = sum(acc_sh)/len(acc_sh)
js_s = sum(js_sh)/len(js_sh)
score["acc"] = acc_s
score["js"] = js_s
print("Different Metrics were: ", score)
return(score)
if __name__ == "__main__":
# to hold loss values after each epoch
loss_all_epochs = []
# to hold different metrics after each epoch
Specificity_ = []
Senstivity_ = []
F1_ = []
acc_ = []
js_ = []
for epoch_i in range(train_epochs):
# training
print(f"Epoch {epoch_i + 1}\n-------------------------------")
t1 = time.time()
loss_i = train(train_loader, model, optimizer, epoch_i, train_epochs)
loss_all_epochs.append(loss_i)
t2 = time.time()
print("It took: ", t2-t1, " unit time")
# metrics calculation on validation data
dummy_list = validate(val_loader, model, epoch_i)
# Add metrics to empty list above
Specificity_.append(dummy_list["Specificity"])
Senstivity_.append(dummy_list["Senstivity"])
F1_.append(dummy_list["F1"])
acc_.append(dummy_list["acc"])
js_.append(dummy_list["js"])
# loss_all_epochs: contains 2d list of tensors with: (epoch, loss tensor)
# converting to 1d list for plotting
loss_1d_list = [item for sublist in loss_all_epochs for item in sublist]
loss_list_numpy = []
for i in range(len(loss_1d_list)):
z = loss_1d_list[i].cpu().detach().numpy()
loss_list_numpy.append(z)
plt.xlabel("Images used in training epochs")
plt.ylabel("Cross Entropy Loss")
plt.plot(loss_list_numpy)
plt.show()
plt.clf()
x = [i for i in range(1, train_epochs + 1)]
# plot 5 metrics: Specificity, Senstivity, F1 Score, Accuracy, Jaccard Score
plt.plot(x,Specificity_, label='Specificity')
plt.plot(x,Senstivity_, label='Senstivity')
plt.plot(x,F1_, label='F1 Score')
plt.plot(x,acc_, label='Accuracy')
plt.plot(x,js_, label='Jaccard Score')
plt.grid(linestyle = '--', linewidth = 0.5)
plt.xlabel("Epochs")
plt.ylabel("Score")
plt.legend()
plt.show()
# tldr: to make layers behave differently during inference (vs training)
model.eval()
with torch.no_grad():
for image_num, (val_images, val_labels) in tqdm(enumerate(val_loader)):
if DISTRIBUTED_TRAINING:
val_images = val_images.cuda()
val_labels = val_labels.cuda()
else:
val_images = val_images.to(config.DEVICE)
val_labels = val_labels.to(config.DEVICE)
# model prediction
val_pred = model(val_images)
# Coverting val_pred from (1, 19, 512, 1024) to (1, 512, 1024)
# considering predictions with highest scores for each pixel among 19 classes
prediction = val_pred.data.max(1)[1].cpu().numpy()
ground_truth = val_labels.data.cpu().numpy()
# replace 100 to change number of images to print.
# 500 % 100 = 5. So, we will get 5 predictions and ground truths
if image_num % 100 == 0:
# Model Prediction
decoded_pred = val_data.decode_segmap(prediction[0])
plt.imshow(decoded_pred)
plt.show()
plt.clf()
# Ground Truth
decode_gt = val_data.decode_segmap(ground_truth[0])
plt.imshow(decode_gt)
plt.show()